Yusuke Oda

Sorry for being late. Could you please provide specific info for this, e.g., environment and reproducible commands?

@Tojens `` is a placeholder: users need to replace it to a specific number of threads.

I thought that the early stopping is usually done based on number of evaluations (e.g. if there is no dev-loss improvement in recent K evaluations, then the trainer would stop)....

Which version of Chainer are you using? Since many internal architectures of Chainer had been dramatically changed by the ver. 1.5 update, basically all code in the "1.4" directory cannot...

This error probably seems some sort of problems in your environment. Please check the CUDA binaries works correctly.

Yes, current implementation calculates BLEU with UNK tokens (if you are using WordVocabulary). This is because of some reasons...we need some additional modules (e.g. tokenizer requested from specific evaluation campaigns)...

After reading your PR, I think it might be more better only replacing UNK ID in ref or hyp to other values which is never used for other word ID...

Adding a switch for UNK treatment is good I think.

Using original texts sometimes cannot calculate correct BLEU when, for example, we are using CharacterVocabulary, since this class does not care of any word separators and it basically could accept...

I am now re-making some data structures to introduce raw tokens for the evaluation in the `new-structure` branch.